Why even use AI? Mocap + features reconstructed from some photo + your acting under that mask + one spam call to steal voice sample + pitch shifter to slightly modify your voice. We have fucking Hololive where vtubers are anime girls streaming in real time. If it means at least $1 mil, you can find the right people to do a good deepfake under a budget. And they won't probably use LLM if they worth employment as it's unreliable and, as you said, needs training data, needs processing power, time etc. You'd be surprised how many things people do as shitposts in AE and Blender. Making Biden say an N-word to Obama is easier than you think when you know who to ask.
It's not an LLM problem, it's their stupid security, verification processes and this person be a gullible idiot.